Anthropic announced it will modify how Claude generates text, embedding steganographic watermarks to comply with an EU regulation requiring all AI-generated text to be marked starting December 2026. The technique works by biasing token selection at inference time — nudging the model toward "green list" words and away from "red list" words at each generation step, creating a statistical pattern detectable only by those holding the secret key. The mechanism is structurally identical to a loaded coin. At each token-generation point, words are sorted into green and red lists deterministically based on a secret key. The model becomes slightly more likely to pick green-list words. Over enough tokens, the pattern becomes statistically detectable. Shorter texts produce lower confidence; longer texts approach certainty. Only Anthropic can detect Claude's watermarks, and Claude cannot detect Gemini's, because each system relies on proprietary keys. John Gruber, the veteran tech blogger, called the system "a perverse adulteration of what it means to write," arguing that constraining word choice necessarily degrades output quality. His core claim: no two synonyms carry identical meaning, so biasing selection away from the optimal word at any decision point makes the writing worse. Steven Murdoch, professor of computer science at University College London, countered that LLMs already make word choices stochastically — the randomness is essential to prevent loops and repetition. "There's going to be no noticeable difference," Murdoch said. Gruber's objection exposes a real tension, but misidentifies where the damage lives. LLMs do not "contemplate" whether "stream" or "brook" better serves a sentence's register. They roll weighted dice. The watermark changes the weighting on those dice. Whether this produces measurably worse prose is an empirical question neither side has answered with data — Anthropic claims "imperceptible" impact, Gruber asserts degradation, and Murdoch calls it a wash. Nobody has published A/B comparisons. The regulation applies to all AI companies operating in the EU, meaning OpenAI, Google, and every other provider will need equivalent systems within months. This creates a new structural layer: provenance detection controlled entirely by the model providers themselves. There is no independent verification mechanism described. Each company holds its own secret key, grades its own homework, and determines what counts as a watermark match. The less-discussed but arguably more consequential justification is model collapse prevention. Training AI on AI-generated content degrades model quality over generations, causing concept confusion. Watermarking enables automated detection and filtering of synthetic text from training corpora — a defensive measure that protects the AI companies' own future models as much as it protects the public. A retracted chemistry paper illustrates the stakes: authors apparently used an AI tool that substituted "mass killing of an ethnic group" for "final solution" in a paragraph about zinc nanogel. Watermarking would have flagged the text as AI-generated before publication. The system serves multiple masters — EU regulators, academic integrity, disinformation detection, and the AI companies' own training pipeline hygiene — but the control architecture concentrates power in the providers, not the public.
